Concerns have been raised about the reliability of medical advice given by artificial intelligence (AI) chatbots. A recent study in Nature Medicine evaluated the accuracy of ChatGPT Health, OpenAI’s free consumer health tool, launched in January 2026, which has millions of users. The researchers compared how well ChatGPT Health handled triage recommendations compared to the best advice from doctors. A total of 960 recommendations were evaluated.

The most dangerous failures happened at the extremes: When conditions were not urgent, 35 percent of ChatGPT Health’s responses recommended more care than necessary, and when conditions were emergencies, 48 percent of the chatbot’s responses failed to recommend urgent or immediate attention. 

When it came to emergencies that are considered “gold standard”, ChatGPT Health did not treat them as seriously as it should have in 52 percent of cases. For example, it sent patients with diabetic ketoacidosis or impending respiratory failure to see a doctor in 24 to 48 hours instead of directing them straight to the emergency room. However, it did a good job with classic emergencies like strokes and anaphylaxis.

The researchers also found that context influenced recommendations. When a family member or friend minimized symptoms, the system was more likely to suggest less urgent care in borderline cases. Crisis support messages for suicidal thoughts also appeared to be inconsistent.